Po-Heng Shen

HorusEye Technology (Subsidiary of Oomii Inc.) · Taiwan

• Designed and deployed a production-grade LLM evaluation infrastructure, reducing evaluation time from 3 hours to 20 minutes (~10×), enabling continuous regression testing for AI systems.
• Optimized AI backend systems (webhooks, CX agents), achieving latency reduction from 4–5s to 0.1–0.2s (~10× improvement) through system redesign — caching strategies, database migration from BigQuery to Cloud SQL.
• Designed and led development of a large-scale AI testing pipeline (>20K LOC), incorporating LLM-as-a-judge and cross-model validation to improve system reliability.
• Owned and delivered core AI components in a 7-figure enterprise project, ensuring stable delivery under tight deadlines and high system load.
• Developed and evaluated Graph RAG systems, improving knowledge retrieval accuracy for complex, real-world datasets.
• Acted as a key technical bridge across teams (PM, backend, design), aligning research-driven AI solutions with production deployment requirements.

IES Dept., University of Kassel · Germany

• Developed and evaluated 6 replay-based continual learning algorithms, reducing catastrophic forgetting to below 5% on 2-year time-series data and establishing 5 new evaluation standards.
• Built and optimized a VAE-based anomaly detection model for the Digital Twin Solar project, boosting F1-Score from 0.96 to 1.00 and cutting detection time from 8.2s to 0.1s on cloud infrastructure.
• Created a VAE model to reconstruct 1-year time-series data (526,000 points) with 96–97% similarity, along with 4 evaluation metrics for quality assessment.

Multicast Digital Integrated Corporation · Taiwan

• Worked on backend debugging, IoT system integration, and network security setup.

Kassel University · Germany  ·  ECTS: 1.9

National Central University · Taiwan  ·  GPA: 3.59 / 4.0